Silently ignore missing arguments in constructor

[t-kalinowski]

As discussed in #445, this PR allows missing arguments in a constructor to "skip" setting the property and avoid invoking a custom setter.

[hadley]

This seems a bit confusing because it is missing?

[t-kalinowski]

The issue happens because person@last_name <- <value> is never set since <value> is missing. Then, when the validator fetches person@last_name, it returns NULL (because attr(x, "does_not_exist") return NULL), which is the wrong type and causes the error.

This whole test probably needs to be rewritten to make sense with the new behavior.

[t-kalinowski]

S7/src/prop.c

Lines 260 to 261 in 15a01a3

	// This is commented out because we currently have no way to distinguish between
	// a prop with a value of NULL, and a prop value that is unset/missing.

[hadley]

If it's not set, should we be using quote(expr = )?

[t-kalinowski]

To do that, we would have to use Rf_setAttrib(object, Rf_install("name"), R_missingArg) in prop<- when the object is first constructed.

quote(expr=) is such a pain to work with at the R level; I'm beginning to think that perhaps we should not create a pattern that requires all object instances to be handled with care, checking with missing() like this.

To make the "deprecate via getter+setter" pattern possible, maybe we should return to what we discussed in #396 (comment) and altogether bypass or not invoke custom setters on initial construction.

We would then need to provide a convenient way to "opt-in" to running the setters without requiring a full custom constructor.

Perhaps we add an argument to new_class(..., initializer = function(self) {}), a function authors can provide. The initializer function will be called with the output of the constructor after the constructor() has returned but before validator() is called. (somewhat analogous to __new__ and __init__ in Python.)

[t-kalinowski]

We may also want to add an initializer arg to new_property, defaulting to setter.

new_property(..., initializer = setter)

[lawremi]

Could we just use class_missing as the sentinel value? Or perhaps something similar but named like arg_missing?

[hadley]

I think we got a bit side-tracked into a much bigger topic than I was thinking. Looking at this with fresh eyes, I think all that I want is for the error message to be:

! <Person> object properties are invalid:
- @last_name must be <character>, not <NULL>

[t-kalinowski]

If/when promise accessors become part of the C "API", the code in this PR should be updated to use them (posit-dev/ark#538 (comment))